atmem/laya-memory-decision-v2

Laya Memory Decision V2 — authentic-conversation research checkpoint

This is a portable multilingual classifier for memory-state decisions. A host memory system supplies a new observation, its authorized retrieved memories, and one bounded Choice, Noul, or Score question. The model returns a typed probability distribution; it does not generate memory text or write a database.

The checkpoint is independent of AtMem. AtMem is one compatible host and owns retrieval, authorization, mutation, provenance, and receipts. Other memory systems can use the same JSON contract in questions.json.

Status

Research-only, uncalibrated checkpoint. It is public for inspection and experimentation, but is not qualified to authorize automatic memory mutation. The calibration split lacks enough relational operations, and the frozen production selector therefore returned blocked.

Data and lineage

Item	Frozen identity
Base model	convaiinnovations/laya-multilingual@1720e3e3357cfe1e281542e223f8273b0890ca34
Laya source	NandhaKishorM/laya@68804629e8ccd9d616d48a40e87de9fabbeae069
Dataset	atmem/stateful-memory-transitions-10pct-v2@92a851b7f1811d3403f8d2282ac1275e6f957922
Dataset manifest	2cb6aac781b07f6734bbbc4a465260850363c77093e2a79c7ded4c98de2c4730
Training run	rlcd-410239-20261013T203312Z
Model SHA-256	17c949572ecd43917ed8d0413c3e2c68252432ef1034623e57caf61d7920ada1
Seed	410239

The dataset contains only human-origin public conversation rows selected from WildChat-1M and LUFY. It contains no generated, synthetic, translated, or hand-authored conversation rows. Transition labels are contextual machine weak supervision produced against frozen scoped SQLite and vector state; they are not human ground truth. Users, conversations, scopes, chronological chains, and duplicate clusters are split together to prevent cross-split leakage.

Typed contract

The host sends a versioned request containing the observation, authorized retrieved memories, canonical generation, and exactly one typed question:

{
  "format": "laya-evaluation-request-v2",
  "schema_version": "2.0.0",
  "request_id": "req_01",
  "scope_id": "opaque-scope",
  "canonical_generation": 184,
  "observation": {"id": "obs_01", "speaker_role": "user", "content": "Actually, call me Sara rather than Mona."},
  "retrieved_memories": [{"id": "mem_41", "content": "The user prefers to be called Mona.", "lifecycle": "active"}],
  "question": {
    "question_id": "formation.operation.v2",
    "type": "choice",
    "instructions": "Select the supported memory-state operation.",
    "options": [
      {"id": "STORE_NEW", "description": "Create distinct durable memory."},
      {"id": "UPDATE_EXISTING", "description": "Add compatible detail."},
      {"id": "SUPERSEDE_EXISTING", "description": "Replace outdated memory."},
      {"id": "DUPLICATE", "description": "Meaning is already represented."},
      {"id": "DO_NOT_STORE", "description": "Do not create memory."},
      {"id": "FORGET_EXISTING", "description": "Retire an authorized target."}
    ]
  }
}

The output is projected to a strict Choice, Noul, or Score response with a full probability distribution. questions.json contains the frozen definitions. A host must preserve complete state, reject truncation, restrict target IDs to the authorized retrieved set, and treat every output as a proposal.

Loading

from laya import Agent

agent = Agent("atmem/laya-memory-decision-v2", device="cpu")
result = agent.system_one(
    {
        "observation": {"id": "obs_01", "content": "Call me Sara."},
        "retrieved_memories": [],
    },
    {
        "formation.is_durable.v2": {
            "type": "noul",
            "statement": "The observation is likely to remain useful later."
        }
    },
    max_len=1024,
    head_max_len=256,
)

Pin the model revision in reproducible or governed deployments.

Training

Training used upstream RLCD/proper scoring for four epochs on one RTX 3090, micro-batch 8, gradient accumulation 8, encoder learning rate 2.5e-5, head learning rate 1e-4, FP32, gradient clipping 1.0, and seed 410239. PyTorch CUDA has no deterministic cumsum kernel used by the RLCD ordered-score reward; the run records a warn-only exception for that operation while retaining fixed seeds and deterministic settings elsewhere.

The predeclared production selector rejected both RLCD and soft-CE because of input truncation, missing rare expected classes, and insufficient calibration power. RLCD was nominated only for research comparison from non-sealed metrics.

Frozen evaluation

The sealed split has 1,999 transitions after privacy exclusions: 1,821 DO_NOT_STORE, 107 STORE_NEW, 61 REVIEW, 8 UPDATE_EXISTING, and 2 DUPLICATE; it contains no SUPERSEDE_EXISTING or FORGET_EXISTING examples.

Metric	Base Laya	Fine-tuned checkpoint
Fully correct transitions	367 / 1,999 (18.36%)	1,475 / 1,999 (73.79%)
Operation accuracy	25.11%	80.54%
Operation macro-F1	9.18%	23.37%
Target accuracy	0.00%	10.00%
Runtime failures	1,269	215
False stores	2	15
Unsafe mutations	10	9
Brier score	0.4957	0.1852
ECE	0.5653	0.4092

The paired fully-correct improvement is +55.43 percentage points; paired bootstrap 95% CI [+53.18, +57.63]. A constant DO_NOT_STORE classifier would score 91.10% aggregate on this imbalanced split, so aggregate accuracy alone is not evidence of a deployable decision model. The fine-tuned checkpoint improves macro-F1 over base Laya but increases false storage and remains poorly calibrated.

Appropriate use

• Research on typed memory-decision contracts and state-aware classification.
• Offline comparison using the published frozen database and transition set.
• Proposal-only integration behind host validation and explicit review.

Do not use this checkpoint as an authorization oracle, autonomous database writer, deletion authority, or calibrated production classifier.

License

Apache-2.0. Source-dataset licenses and provenance are recorded per row in the dataset card and manifest.